This rewrites replaces a batched matmul by a core matmul by raveling the batched dimensions of the batched input, doing a 2D matmul and then unravelling the batched dimensions of the output.

This sidesteps the Blockwise Dot operation and allows specialization into BLAS routines.

The idea was taken from these two discussions:
https://github.com/numpy/numpy/issues/7569
https://github.com/numpy/numpy/issues/8957